Emily (Miller) Dorne

So that we can track breaking changes in package version updates, we should set up a workflow to run the tests weekly and file an issue for failed tests.

Brew now installs ffmpeg 5, which causes at least some tests to fail (notably the CLI test that predicts on an actual video). More work is needed to understand what...

Currently, model predictions will always get written to a file called `zamba_predictions.csv` (in the specified save_dir). We may want to add an `output_csv_name` arg to parameterize this. Relatedly, we should...

Expose ability to set learning rate for models

4

Right now, the only options is to use `auto_lr_find` or not. If that is False, default learning rate is 0.001. We should let users specify this learning rate. `ZambaVideoClassificationLightningModule` has...

Create demo of v2

1

Consider using: https://asciinema.org/ which allow users to copy/paste code from video

We should have one canonical place for configurations in the docs. Currently, configuration params are documented multiple times and this means it can be easy for the docs to conflict...

Rename zamba models

1

We're moving toward a world where we only have one model architecture (time distributed) and various geographies. Should we stop calling the model `time_distributed` and start naming based on African...

Uncouple multiprocessing settings

1

Right now, we set the `multiprocessing_context` for the Trainer based on the `num_workers` used for the data loader https://github.com/drivendataorg/zamba/blob/master/zamba/pytorch_lightning/utils.py#L67-L71 https://github.com/drivendataorg/zamba/blob/master/zamba/models/model_manager.py#L283-L286 It would be good to separate those out for a...

It would be helpful to provide guidance on recommended compute set ups for running zamba and the corresponding time estimates for how long inference will take. If I have 1,000...

[`auto_lr_find` does not support DDP](https://pytorch-lightning.readthedocs.io/en/stable/advanced/training_tricks.html#learning-rate-finder). We [set the multiprocessing context](https://github.com/drivendataorg/zamba/blob/master/zamba/pytorch_lightning/utils.py#L67-L71) based on the number of workers. This leads to a scenario where a user cannot use `auto_lr_find` with multiple workers,...